Performance and testing of threshold identification methodsFunded under: FP7-JRC
We have applied 3 statistical methods for threshold detection in the form of abrupt changes in the mean value of regular time series. The methods differ in their scope and their user skill requirements: one works as an add-in module for Microsoft Excel, while the other two require some familiarity with the R environment for statistical computing to be useful. The 3 test data sets differ in the extent to which the existence of a threshold is known or visually self-evident. The sequential t-test method of Rodionov (2004) seems to give results consistent with the other methods on time series low short-term variability, such as annual average data with low year-to-year variability. But it gives erratic and inconsistent results on time series with higher frequency content, such as noisy data or data with strong seasonal variability. Both the sequential F-tests and Empirical Fluctuation Process (EFP) test methods (Zeileis et al 2003) seem to give consistent results on both seasonal and annual data, but the EFP method appears to be favour for testing the initial hypothesis of whether a time series has one or more change-points, or not. The sequential F-test is recommended for further exploration of change-point location (and corresponding confidence limits) if the EFP test is significant.
Bibliographic Reference: EUR 22647 EN (2007), 22 pp. Free of charge
Availability: Katalogue Number: LB-NA-22647-EN-N The PDF version can be downloaded from: http://bookshop.europa.eu
Record Number: 200719430 / Last updated on: 2007-09-18
Original language: en
Available languages: en